Advancing LOINC Mapping with AI: Insights from LARA’s RAG Implementation and Expert Evaluation

Accurate mapping of laboratory codes is crucial for achieving semantic interoperability in Electronic Health Records (EHRs), yet the idiosyncrasy of local codes and names poses a significant challenge. This task is time-consuming and compounded by the absence of support tools for automatic mapping, especially for the Italian language. As well as being required by current Italian legislation, fostering LOINC coding adoption is pivotal for two primary reasons: a) laboratory reports are the most frequently indexed clinical documents in the patients’ EHRs, and b) standardized data can enhance public health monitoring and planning. This paper presents the development of the LOINC AI-based Mapping and Remapping Assistant (LARA), focusing on the implementation of Retrieval-Augmented Generation (RAG) to automate the mapping of local laboratory test codes to LOINC. LARA leverages a combination of text similarity, domain-specific embeddings, and generative AI to enhance the accuracy and efficiency of LOINC mapping. The study also evaluates LARA’s performance based on human expert feedback, analyzing the system’s ability to suggest accurate LOINC codes for laboratory test descriptions. The findings highlight LARA’s strengths in handling synonymy and mapping specificity while identifying areas requiring further refinement, such as handling ambiguous test descriptions and linguistic variations. Expert insights provide valuable directions for improving LARA’s interpretative capabilities. The results demonstrate LARA’s potential in reducing the manual effort required for LOINC mapping and reliability in the identification of the correct code to represent equivalent semantic concepts. AI-based supporting tools could thus foster the adoption of standardized coding systems.